Awesome

SSM Benchmark on Long-range Arena (LRA)

Note: This repo is still under development. Currently only the configs for CIFAR-10 are available, but we will add the other configs including links to some example W&B runs in the coming days.

This repository provides the implementation for the paper:

State Space Models as Foundation Models: A Control Theoretic Overview
Carmen Amo Alonso*, Jerome Sieber*, Melanie N. Zeilinger
Submitted to the 63rd IEEE Conference on Decision and Control (CDC), 2024.
(*equal contribution)
arXiv

References

This repository reuses code from various other repositories.

• mamba-1.1.4/ contains the code of mamba-ssm (release 1.1.4). We include it here to facilitate on-the-fly changes to the original Mamba architecture.
• accelerated-scan-0.1.2/ contains the code of accelerated-scan (release 0.1.2), which is used to implement the Hawk models.
• The Hawk models are adapted versions of the implementation provided by Hippogriff.
• dataloaders/ is an adapted implementation of the code provided in the S5 repository.

Requirements & Installation

To run the code on your own machine, run pip install -r requirements.txt. Then, install mamba-ssm in development mode with pip install mamba-1.1.4/. Finally, build the accelerated-scan package by running python -m build in the accelerated-scan-0.1.2/ folder and install it with pip install accelerated-scan-0.1.2/.
(Warning: The accelerated-scan version shipped with this repo only requires torch >= 2.0.0, i.e., the triton method will not work properly.)

Data Preparation

See the dataloaders README for more details.

Example Usage

The configurations to run the LRA experiments from the paper are located in configs/. For example, to train Mamba on the LRA text task (character level IMDB), run python train.py --config imdb-mamba.yaml. To log with W&B, fill in the wandb arguments in the config files.

Repository Structure

Directories and files that ship with GitHub repo:

accelerated-scan-0.1.2/ Snapshot of accelerated-scan release 0.1.2, with downgraded torch requirement.
configs/                YAML configuration files for each experiment.
dataloaders/            Code mainly derived from S5 processing each LRA dataset.
mamba-1.1.4/            Snapshot of mamba-ssm release 1.1.4. If installed in development mode, change the model in here.
models/                 Model definitions of Mamba & Hawk.
requirements.txt        Requirements for running in GPU mode (installation can be highly system-dependent).
train.py                Training loop entrypoint.

Directories that may be created on-the-fly:

data/                   Default data path used by dataloaders.
wandb/                  Local WandB log files.

Citation

Please use the following BibTex entry when citing our work:

@misc{alonso2024state,
      title={State Space Models as Foundation Models: A Control Theoretic Overview}, 
      author={Carmen Amo Alonso and Jerome Sieber and Melanie N. Zeilinger},
      year={2024},
      eprint={2403.16899},
      archivePrefix={arXiv},
      primaryClass={eess.SY},
      url={https://arxiv.org/abs/2403.16899}
}

Please reach out if you have any questions.